Machine Learning Potential
A method that fits interatomic potential energy surfaces and forces using machine learning models, with first-principles calculation data as training labels. Representative work includes Behler-Parrinello symmetry functions, GAP, and neural network potentials such as ANI and MACE.
Timeline (last 90 days)
Machine learning potentials (MLP) are methods that use first-principles calculation data as training labels and fit interatomic potential energy surfaces and forces with machine learning models, with representative work including Behler-Parrinello symmetry functions, Gaussian approximation potentials (GAP) and neural network potentials (such as ANI and MACE).
Traditional machine learning potentials and graph neural networks are usually trained for specific tasks, have limited generalization ability, and have limited extrapolation ability for structures outside the training distribution.
Molecular dynamics combined with machine learning potentials is currently a common means of handling finite-temperature property problems, but its computational cost is far higher than single-point property prediction.
All Facts (3)
Machine learning potentials (MLP) are methods that use first-principles calculation data as training labels and fit interatomic potential energy surfaces and forces with machine learning models, with representative work including Behler-Parrinello symmetry functions, Gaussian approximation potentials (GAP) and neural network potentials (such as ANI and MACE).
50%UnverifiedTraditional machine learning potentials and graph neural networks are usually trained for specific tasks, have limited generalization ability, and have limited extrapolation ability for structures outside the training distribution.
50%UnverifiedMolecular dynamics combined with machine learning potentials is currently a common means of handling finite-temperature property problems, but its computational cost is far higher than single-point property prediction.
50%